Laurent Lessard

Associate Professor of MIE @Northeastern. Interested in control theory, optimization, math, and puzzles. The glass is twice as large as it needs to be.

[Laurent Lessard @LaurentLessard]

My contribution to the #Wordle discourse, with help from @VincentTjeng and @evanrsparks.

A definitive look at the game: insights, heuristics, optimal strategies, and the limits of what is possible.

https://t.co/jSRXV1BYr5

[Laurent Lessard @LaurentLessard]

"Deadlines are final and no extensions will be given"

You got me @IEEECDC2025 --- I totally fell for it! Happy April Fools everybody!

[Krishnamurthy (Dj) Dvijotham @DjDvij]

Can your AI keep up with dynamic attackers? In a paper to appear at #AISTATS2025 with @avibose22 @LaurentLessard and Maryam Fazel, we study robustness to learning algorithms to dynamic data poisoning attacks that can adapt attacks while observing the progress of learning

[Laurent Lessard @LaurentLessard]

@gravity_levity It doesn't decay quite that quickly, but it's still fast:
P(1) ~ 1/k!
So basically: -log(P(1)) = O(k*log(k))
rather than: -log(P(1)) = O(k^2),
which is what you had conjectured.

[Laurent Lessard @LaurentLessard]

This week's #Fiddler puzzle is a random walk: From (k), you hop to (k-1) with probability 1/k and hop to (k+1) otherwise. If you start at (2), what is the probability you eventually hop to (1)?

I wrote a short tutorial here:
https://t.co/V0QhIIEiLw

[Laurent Lessard @LaurentLessard]

This week's #fiddler: Assign distinct prime numbers to the vertices of a dodecahedron so the sum of the vertices of each face sum to 2025. I modeled this as an integer linear program. Here is one possible solution. @xaqwg

My write-up: https://t.co/vFbpDurcNP

[Laurent Lessard @LaurentLessard]

Update: after 2 hours and 35 minutes on the tarmac, we are finally deplaning.

[Laurent Lessard @LaurentLessard]

Hey @AerLingus I'm on a flight from Dublin to Boston and we've been stuck on the tarmac at @loganairports after landing for TWO HOURS (and counting). Apparently other planes are being de-iced which is preventing us from deplaning...

[Laurent Lessard @LaurentLessard]

The pilot keeps saying there is "nothing he can do". You mean to tell me there are no other gates available at the airport? We can't deplane on the tarmac? This is a full flight with 8 seats across and we've been on this plane for almost 10 hours now.

[Laurent Lessard @LaurentLessard]

This week's #Fiddler: what is the minimum-energy way to pack 9 particles in a square box, where energy is measured as the sum of 1/dist over all pairs of particles?

Turns out it's NOT a regular 3x3 lattice! @xaqwg

write-up:
https://t.co/tBfK0Jy2yq

[Laurent Lessard @LaurentLessard]

Halloween #fiddler puzzle: A bag contains N Reese's cups and an unknown number of candy corn pieces. You reach into the bag k ≤ n times at random and pull out a Reese's each time! How many candy corn do you expect to be in the bag? @xaqwg

My write-up:
https://t.co/z2DxoGC9xr

[Lars Lindemann @LarsLindemann2]

Great pleasure of hosting @LaurentLessard at USC who spoke about "An automatic system to detect equivalence between iterative algorithms" 🙂

[Laurent Lessard @LaurentLessard]

@SpottedGeck @xaqwg Nice! You can approximate further using:
Φ(x)-Φ(-x) ≈ 2x*Φ'(0)
and you get:
p(n) ≈ √6 / √(πn)

[Laurent Lessard @LaurentLessard]

@tom_keith1 And I believe this is the correct answer also! I derived it in my blog post https://t.co/uR5jpaaCQv

[Laurent Lessard @LaurentLessard]

This week's #Fiddler by @xaqwg: if x1,...,xn are uniform on [0,1], find the probability that
round(x1)+...+round(xn) = round(x1+...+xn)

It's not so simple! Here's a visualization of n=4, (hypercube slices), for which the answer is 115/192

My write-up:
https://t.co/BPiDlhtvY0

[Laurent Lessard @LaurentLessard]

@tom_keith1 Relevant: https://t.co/Y3EnU3gpdc
This is the sequence of numerators, i.e. 2^n*n!*p(n)

[Laurent Lessard @LaurentLessard]

Thanks for reading! Let us know if you have any comments/questions!

[Laurent Lessard @LaurentLessard]

New preprint with Joao Cavalcanti and @Ashia__Wilson!
https://t.co/gkaAWYvoBd

tl;dr: If you make your backtracking line search routine ADAPTIVE, you get faster continuous optimization algorithms with no additional computational burden!

Longer summary below!🧵

[Laurent Lessard @LaurentLessard]

One such example (we discuss others in the paper) is quasi-Newton methods. For these methods, line search serves the purpose of guaranteeing global convergence rather than finding a "best" step size. Once we enter a local convergence regime, adjustments are no longer beneficial.